perf(context): stop redoing per-query work per candidate on long prompts

[danusha2345]

Refs #2184, follow-up to #2190.

#2190 measured codegraph prompt-hook growing faster than prompt length and noted three hot spots that redo work whose answer cannot change within one retrieval. This PR removes those three and nothing else: ranking is untouched and outputs are byte-identical.

Change

• Path relevance. scorePathRelevance re-split the whole query and re-ran extractSearchTerms for every candidate path (explore's CamelCase/compound steps score up to ~400 LIKE hits per query symbol against the full prompt). The query half is now prepared once (preparePathRelevanceQuery) and each path scored against it (scorePreparedPathRelevance); scorePathRelevance keeps its signature and delegates. searchNodes prepares once per call instead of once per result.
• deprioritize config. The predicate stat'ed codegraph.json per candidate. QueryBuilder.setDeprioritizedPathMatcher now takes a source that returns the predicate for the config in force; searchNodes and ContextBuilder.findRelevantContext take one per pass. Edits to codegraph.json are still picked up by the next search (the explore/query relevance: a generic token's exact name-match overboosts in peripheral dirs — follow-up to #746 #982 "config written after open" test passes unchanged).
• getAllNodeNames. The fuzzy fallback re-read SELECT DISTINCT name FROM nodes once per query term that FTS/LIKE miss. It is now memoized per database change stamp, exactly like getDominantFile (Performance: getDominantFile adds ~5s to every generic explore on large repos #1864): not inside a transaction, dropped on rebind(). The list is held through a WeakRef, so it survives one retrieval but a large index's name list is not pinned in a long-lived server.

Note: QueryBuilder is exported, so two signatures change for library users: setDeprioritizedPathMatcher takes a factory, and getAllNodeNames() returns readonly string[] (the list is shared). If you'd rather keep the old setter, I can add a separate one instead.

Not in this PR

The remaining cost is linear in the number of query terms/symbols (per-term searchNodes, per-symbol substring scans). Capping or deduplicating them changes ranking, so it is left for a separate decision.

Verification

• Identity: on a 377-file index (codegraph's own src/ + ui/src), codegraph_explore output, buildContext markdown and the findRelevantContext subgraph are byte-identical to main for 18 prompts (short symbol queries; 500 B–8 KB prose; 500 B–8 KB subagent reports), with and without a deprioritize config and a project name. searchNodes id+score lists match on 16 terms including fuzzy typos. The prompt-hook's own stdout matched on all 16 timed prompts.
• New test __tests__/node-names-cache.test.ts: a 14-term query that falls through to fuzzy reads the name list once (14 times on main), and a sync() is seen.
• Full suite on Linux passes.
• Prompt-hook wall time (median of 3, includes ~0.4 s process start):

prompt	main	this PR
2 KB prose	1.8 s	1.1 s
4 KB prose	2.9 s	1.5 s
8 KB prose	5.7 s	2.4 s
2 KB subagent report	2.2 s	1.2 s
4 KB subagent report	4.6 s	2.2 s
8 KB subagent report	8.6 s	3.0 s

Short symbol queries are unchanged (~0.4 s).

🤖 Generated with Claude Code